Hyperpolarized Magnetic Resonance and Artificial Intelligence: Frontiers of Imaging in Pancreatic Cancer

Background There is an unmet need for noninvasive imaging markers that can help identify the aggressive subtype(s) of pancreatic ductal adenocarcinoma (PDAC) at diagnosis and at an earlier time point, and evaluate the efficacy of therapy prior to tumor reduction. In the past few years, there have been two major developments with potential for a significant impact in establishing imaging biomarkers for PDAC and pancreatic cancer premalignancy: (1) hyperpolarized metabolic (HP)-magnetic resonance (MR), which increases the sensitivity of conventional MR by over 10,000-fold, enabling real-time metabolic measurements; and (2) applications of artificial intelligence (AI). Objective Our objective of this review was to discuss these two exciting but independent developments (HP-MR and AI) in the realm of PDAC imaging and detection from the available literature to date. Methods A systematic review following the PRISMA extension for Scoping Reviews (PRISMA-ScR) guidelines was performed. Studies addressing the utilization of HP-MR and/or AI for early detection, assessment of aggressiveness, and interrogating the early efficacy of therapy in patients with PDAC cited in recent clinical guidelines were extracted from the PubMed and Google Scholar databases. The studies were reviewed following predefined exclusion and inclusion criteria, and grouped based on the utilization of HP-MR and/or AI in PDAC diagnosis. Results Part of the goal of this review was to highlight the knowledge gap of early detection in pancreatic cancer by any imaging modality, and to emphasize how AI and HP-MR can address this critical gap. We reviewed every paper published on HP-MR applications in PDAC, including six preclinical studies and one clinical trial. We also reviewed several HP-MR–related articles describing new probes with many functional applications in PDAC. On the AI side, we reviewed all existing papers that met our inclusion criteria on AI applications for evaluating computed tomography (CT) and MR images in PDAC. With the emergence of AI and its unique capability to learn across multimodal data, along with sensitive metabolic imaging using HP-MR, this knowledge gap in PDAC can be adequately addressed. CT is an accessible and widespread imaging modality worldwide as it is affordable; because of this reason alone, most of the data discussed are based on CT imaging datasets. Although there were relatively few MR-related papers included in this review, we believe that with rapid adoption of MR imaging and HP-MR, more clinical data on pancreatic cancer imaging will be available in the near future. Conclusions Integration of AI, HP-MR, and multimodal imaging information in pancreatic cancer may lead to the development of real-time biomarkers of early detection, assessing aggressiveness, and interrogating early efficacy of therapy in PDAC.


Introduction
There is an unmet need for noninvasive surrogate markers that can help to identify the aggressive subtype(s) in a pancreatic lesion at an early time point [1]. In contrast to the declines in cancer-related deaths from other malignancies, progress in the management of pancreatic ductal adenocarcinoma (PDAC) has been slow, and the incidence of cancer-related deaths due to PDAC continues to rise [2]. PDAC develops relatively symptom-free, and is one of the leading causes of cancer-related deaths in the United States. In 2020 alone, it was estimated that approximately 57,600 people (30,400 men and 27,200 women) would be diagnosed with PDAC, and approximately 47,050 people (24,640 men and 22,410 women) were projected to die of the disease [3]. Early detection of PDAC is unusual and typically incidental, with the majority (~85%) presenting with locally advanced or metastatic disease when surgery, the only curative modality, is not an option. Overall, PDAC is associated with a dire prognosis and a 5-year survival rate of only 8% [3]. The absence of early symptoms and lack of a reliable screening test have created a critical need for identifying and developing new noninvasive biomarkers for the early detection of PDAC [1].
Hyperpolarization (HP)-based magnetic resonance (MR) has become a major new imaging modality by providing valuable information on previously inaccessible aspects of biological processes owing to its ability for detecting endogenous, nontoxic 13 C-labeled probes that can monitor enzymatic conversions through key biochemical pathways [4][5][6]. Clinical trials with this modality are ongoing at several centers worldwide [7]. HP-MR provides an exciting opportunity to identify and understand early metabolic aberrations, enabling the detection of advanced pancreatic preneoplastic lesions and PDAC at the smallest size for which no methods of detection currently exist. In general, cancer, and PDAC in particular, is considered a paradigm of genetically defined metabolic abnormalities. Genetic mutations can trigger specific signaling pathways that are associated with metabolic transformations, which can potentially be detected by HP methods with a high degree of sensitivity.
In conventional MR, the signal measured is generated from the abundance of hydrogen in the body, specifically water [8].
Organic molecules at high concentration in the body with a high abundance of hydrogens such as choline, lipids, and lactate can also be measured using MR. Other nuclei such 13 C and 15 N can also be measured using MR, but their utility in living systems is low due to their low abundance in nature (the natural abundance of 13 C is 1%) and their smaller gyromagnetic ratio compared to that of hydrogen [9]. HP enables these nuclei to be observed in vivo.
HP allows for >10,000-fold sensitivity enhancement relative to conventional MR, and is a nontoxic, nonradioactive method for assessing tissue metabolism and other physiological properties [10][11][12][13]. There are four established methods for producing HP probes: (i) dynamic nuclear polarization (DNP) [10,13], (ii) optical pumping of noble gases [14], (iii) the brute force approach [15], and (iv) parahydrogen-induced polarization [16]. The detailed physics of these HP methods can be found elsewhere [17]. The most common and widely used method for HP is DNP, in which magnetization is transferred from the unpaired electrons (usually from added radicals) to the isotopically labeled probe [17]. This transfer of magnetization occurs under microwave irradiation at a low temperature of 1.5 K and a high magnetic field of 3 T. Development of the dissolution DNP technique in 2003 [4] opened a new avenue to monitor in vivo metabolism, enabling the detection and tracking of the fate of metabolites containing low-abundance nuclei such as 13 C [18]. The routine dissolution DNP instrument employed, which carries out HP in the preclinical setting, is HyperSense (Oxford Instruments, UK), as shown in Figure S1 in Multimedia Appendix 1. A clinical polarizer is available for performing real-time metabolic profiling in humans (SPINLab, GE Healthcare) and over 20 such polarizers have been installed worldwide [19].
The most commonly used HP probes to track the pathways of interest are 13 C-enriched probes, which are either uniformly or selectively enriched. The other reason to employ 13 C-enriched molecules is the comparatively longer longitudinal relaxation time (T 1 ) of the 13 C nucleus compared to that of other nuclei.
The high 13 C signal of HP probes and the fact that an HP signal is carried over in the products of biochemical transformation allow investigators to interrogate biochemical reactions in real time. These probes are usually part of essential biochemical reactions such as glycolysis (glucose and pyruvate) and the tricarboxylic acid (TCA) cycle (succinate, fumarate, and glutamine).
HP-MR experiments have been performed mostly in preclinical models to date, and HP-MR is not currently routinely used in clinical settings. However, several clinical trials have been performed or are ongoing [5]. HP-MR in the preclinical setting involves injecting the HP probe dissolved in a biocompatible solvent into the tail vein of rodents. The probe diffuses through the blood to populate in well-perfused body tissues. After entering the extracellular fluid, the molecule is taken up into the cells with the help of endogenous transporters. All of these processes must occur before the HP signal decays, which is determined by the decay time (ie, T 1 ) of the HP probe. For most probes, T 1 ranges from 15-20 seconds to approximately 1 minute. Hence, it is important to dissolve the probe in the solvent immediately and inject into the animals quickly to avoid loss of the HP signal due to relaxation. A specially designed proton volume coil and 13 C surface coil are used to receive the signal from the enriched HP 13 C probe in vivo.
The utility of HP-MR is not only simply tracking the probe diffusing inside the body but also its ability to visualize downstream metabolic products of injected probes converted by endogenous enzymes [5]. HP-MR can be used to quantify in vivo metabolic flux in real time. However, all processes must be completed within the time frame of T 1 of the HP probe. Therefore, only relatively fast biochemical reactions can be visualized.
Glycolysis (the breakdown of glucose) is a multistep process that eventually yields pyruvate in the cytosol. Pyruvate is the final breakdown product of glucose in glycolysis and is preferably converted to lactate. The high dependence of cancer cells on glucose and glycolysis is often referred to as the Warburg effect after the initial discovery of this dependence by Dr. Otto Warburg [20]. Therefore, HP [1-13 C]-pyruvate is the most common HP probe for determining glycolytic flux in cancer. Another key point is that pyruvate is taken up rapidly by monocarboxylate transporters [21]. In the cytosol, the HP pyruvate has four important fates [22]: (i) conversion to lactate; (ii) conversion to alanine; (iii) transport into the mitochondria and conversion to carbon dioxide; and (iv) conversion to acetyl-coenzyme A to be utilized in the TCA cycle, which can be tracked by labeling the first carbon of pyruvate ( Figure 1a). When HP-pyruvate is injected into an animal, the signal is recorded from an anatomical imaging slice placed in the tissue of interest. An example of a metabolic HP-MR spectrum is shown in Figure 1b. The flux from pyruvate to a downstream metabolite can be visualized and evaluated using either TopSpin (Bruker BioSpin GmbH) or MestReNova (Mestrelab Research) in either of the two following ways: by measuring the ratio of signals integrated over time (eg, lactate-to-pyruvate ratio, alanine-to-lactate ratio) [23] or by calculating the Kp value (according to the Bolch equation): K PL (pyruvate to lactate) and K PA (pyruvate to alanine) [24]. (b) Downstream products of pyruvate metabolism such as lactate and alanine can be imaged using hyperpolarized magnetic resonance. A 3D, real-time readout of the signals, as shown here, can be created using standard software such as Chenomx.
In summary, HP-MR provides a unique opportunity to measure real-time metabolic signals arising in the tissue of interest with over 10,000-fold sensitivity enhancement that cannot be interrogated by other imaging techniques. The provided outcome is the spectroscopic signatures of the metabolites of interest that are recorded as resonances at different and unique chemical shifts (Figure 1b). The HP-pyruvate signal undergoes decay once it is hyperpolarized with a characteristic decay constant (T 1~5 0 seconds) as well as the downstream products of the metabolism (eg, alanine and lactate). Overall, there is a time window of 3×T 1 (~150 seconds) to accomplish this real-time metabolic imaging, and this short time frame is a major limitation of HP-MR. Fast MR sequence design along with powerful and rapid imaging gradients can help in acquiring more sensitive and informative spectra in the future to mitigate this limitation. Several MR imaging (MRI) companies such as GE Healthcare, Siemens, and Bruker have devoted considerable research investment on this matter.
Artificial intelligence (AI) is a fast-developing research field in which machines are utilized to learn from observations to mimic human intelligence. Kaplan et al [25] define AI as a system's ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation. Over the last decade, deep learning has dramatically reshaped AI research. With the development of deep learning, a subfield of AI, and recognition of its potential in feature extraction and flexibility, it has increasingly been applied to numerous medical scenarios such as diagnosis, health care delivery optimization, genomics, and drug discovery [26][27][28][29][30][31]. Machine learning has been utilized for online health care management [32], disease prevention [33], clinical note processing [34], and management of chronic diseases [35]. AI has been leveraged for diagnosis and localization of regions of interest (ROIs) using a vast array of medical images such as optical images, MRI, X-rays, and computed tomography (CT) [36][37][38][39][40][41]. As a result, there is a great opportunity to utilize AI for the early detection of cancer such as PDAC.
Deep-learning algorithms rely on neural networks, which mimic the process of information transformation by neurons in the biological brain [42]. Neural networks adaptively learn features from observations during training and translate the input data to high-dimensional representations suitable for classification or regression tasks. The success of deep-learning algorithms is rooted in their multiple stacked layers and efficient feature extraction, often explained as a powerful representation learning method. Each layer consists of multiple neurons transforming the information nonlinearly by an activation function. This architecture allows for high-level interactions between transformed features coming from the previous layers to contribute to the output. Hence, deep-learning algorithms could automatically optimize the parameters and learn a high-level representation of input data aligned with the target task.
As shown in Figure 2, we believe that the knowledge gap of "early diagnosis of pancreatic cancer with noninvasive imaging" is an elephant in the dark that cannot be accomplished with a single modality. Pancreatic cancer at the very early stages is completely asymptomatic. Conventional anatomical imaging cannot detect any of these early stages of premalignancy of this deadly disease when therapeutic or early surgical interventions can be most effective. Conventional MRI can detect intraductal papillary mucinous neoplasms (IPMNs) where epithelial pancreatic cystic tumors of mucin-producing cells arise from the pancreatic ducts [43]. Although IPMNs are benign tumors, they can progress to pancreatic cancer in some cases [43]. However, MRI as well other imaging modalities fail to detect any other premalignant lesions such as pancreatic intraepithelial neoplasia (PanIN), which is a more commonly accepted mechanistic pathway of the tumorigenesis of PDAC [44]. It is important to recognize that an individual with even stage I (localized) pancreatic cancer has a 5-year survival rate of only 39% [45]. This emphasizes the point that early detection in pancreatic cancer must occur at stages earlier than clinical stage I. HP-MR can detect metabolic changes at very early stages of lesion formation in the pancreas; however, this is more of an MR spectroscopic technique than an MRI modality. Moreover, the signal from HP compounds lasts no more than a few minutes that allow for a rapid acquisition of dynamic metabolic flux measurements in the organ of interest. This review will focus on the introduction of AI approaches to CT and MRI datasets, and the applications of HP-MR in pancreatic cancer. In the Results section, we summarize the strengths and weaknesses of each technique, and discuss our solution to leverage the unique strengths of AI to learn biomarkers from both HP-MR and MRI modalities, in addition to the available pathology and immunohistochemistry data to bridge this crucial knowledge gap. Our laboratories are currently pursuing an AI approach using an HP-MR dataset as applied to PDAC, the results of which will be published in the near future. In addition, we discuss the broad range of HP probes used to interrogate physiological functions such as metabolism and pH, which may expand the scope of applying AI to the functional imaging of PDAC.

Methods
A systematic review was performed following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) extension for scoping reviews (PRISMA-ScR) guidelines. Studies addressing the utilization of HP-MR and/or AI for early detection, assessment of aggressiveness, and interrogation of the early efficacy of therapy in patients with PDAC cited in recent clinical guidelines were extracted from the PubMed and Google Scholar databases. The studies were reviewed following predefined exclusion and inclusion criteria, which were grouped based on the utilization of HP-MR and AI in PDAC diagnosis.
Application of the HP-MR technique in pancreatic cancer is still nascent. We have reviewed every paper published in this broad area up to November 2020. Taken together, we have summarized our review in two tables. Table 1 summarizes all 13 C-labeled HP probes employed in interrogating different metabolic pathways in pancreatic cancer systems, and Table 2 summarizes all published applications of HP-MR in preclinical models of PDAC. In all, we have classified all of the physiological applications of HP-MR in pancreatic cancer under seven categories. The details of the deep-learning methods and HP-MR in different PDAC applications are discussed in the Introduction section above and in the relevant subsections of the Results.

Characteristics of Retrieved Articles
For AI applications in pancreatic cancer, we retrieved 112 articles from the two sources, including 87 articles from PubMed and 25 articles from Google Scholar. An article was included if it satisfied our inclusion criteria: (1) written in English; (2) utilized AI/machine learning/deep learning for prediction, diagnosis, or classification; and (3) proposed a novel method of employing AI for PDAC ( Figure 3). Review, evaluation, and comparison papers were therefore not included. Among the retrieved papers, a total of 17 met the inclusion criteria ( Figure  3, Table 2, and Table S1 in Multimedia Appendix 1). The selected papers were grouped into six categories based on how AI was utilized in the context of PDAC to recognize the gaps in the previous studies and to discuss the novel approaches that fill the current gaps in detecting PDAC by imaging modalities. For HP-MR, we retrieved and reviewed all papers published in this broad area up to November 2020, which included six preclinical studies and one clinical study. We also reviewed several HP-MR-related articles (52 articles) that described new probes that can be applied in many functional future applications in PDAC. These references are not included in the PRISMA flow chart in Figure 3, as they have not yet been demonstrated in PDAC imaging and spectroscopic applications.

Context for Application of HP-MR in Pancreatic Cancer
PDAC tumors can be removed by surgery if detected early [23]. There is unequivocal evidence that diagnosis of PDAC at earlier, resectable stages has a profoundly favorable impact on prognosis [ to 30%-60% for tumors <2 cm, and as high as 75% for minute lesions under 10 mm in size [46,47]. Unfortunately, most tumors are diagnosed at a late stage, once advanced into the local blood vessels and other body organs, and can no longer be excised. Thus, there is an urgent call to develop noninvasive imaging modalities for the early detection of PDAC, especially in high-risk patients (eg, those with a familial predisposition, long-standing diabetes, or chronic pancreatitis) [48]. Unlike other cancers such as breast or prostate cancer that have close to 100% survival if detected at early stages, PDAC is associated with a survival rate of only 39% even when detected at stage I [45]. Therefore, there is an urgency to develop novel methods for the detection of preneoplastic lesions in the pancreas.

Grading of PDAC
The type of treatment administered is often dependent on the tumor grade; therefore, there is a need for noninvasive methods to determine tumor grade. HP-MR uses metabolic changes to determine a grade [49]. Inside a PDAC tumor, the malignant cells become dependent on glycolysis for energy generation (Warburg effect). Dutta et al [50] recently reported that the aggressiveness of PDAC is directly correlated to pyruvate-to-lactate conversion measured using HP-MR (  [63]. These studies reveal the potential for the conversion of HP-pyruvate to lactate in the early detection of PDAC. In addition, the HP pyruvate-to-lactate ratio may be used for staging tumors in the context of their aggression, although how this paradigm would fit in with the existing standards of staging is debatable (stage I or II: surgically resectable; stage III: locally advanced, unresectable; stage IV: metastatic) [48]. A very promising use of pyruvate-to-lactate flux is to identify PDAC advancing toward stage IV (metastasis) because these tumors show higher pyruvate-to-lactate conversion compared to that of less aggressive pancreatic cancer [50].

Early Assessment of Treatment Response
One of the promising utilities of HP-MR is its ability to assess treatment response early during the regimen; this has been established for solid tumors characterized by "aggression correlated with increased glycolysis." This technique can thus complement the standard fluorodeoxyglucose-positron emission tomography imaging, which can only detect changes in tumor size (rather than intracellular metabolic changes) once it shrinks in response to a long-term regimen of chemotherapy or radiation therapy. Table 2 summarizes four published studies that show how HP-MR can be employed to predict responders (prognostic biomarkers) or assess treatment response early in PDAC tumors or cells [54,[63][64][65]. The treatment efficacy of drugs (hypoxia-activated prodrugs, β-lapachone, and LDH-A inhibitors) evaluated using HP [1-13 C] pyruvate has only been studied in preclinical models to date; however, the preclinical data illustrate the ability of HP-MR to assist clinical trials by providing a framework for personalized medicine. HP-MR can provide information about the efficacy of drugs at an early stage that can lead to changes in clinical management, enabling the clinician to change the drug for a nonresponding patient to a more effective drug at an early stage.
Wojtkowiak et al [64] (  [65]. LDH-A converts pyruvate to lactate in the presence of its cofactor nicotinamide adenine dinucleotide hydrogen (NADH). Inhibition of LDH-A is a metabolic vulnerability that can be exploited for cancer treatment, and hence FX11 was evaluated in PDAC animal models. The drug was injected once daily for 4 weeks using PDX mouse models with tumors in their flank. The drug efficacy was tested using HP [1-13 C] pyruvate, which was injected into the mice prior to the start of treatment and 7 days after treatment, prior to any changes in tumor volume. Mice responding to the treatment showed a decreased lactate-to-pyruvate ratio after FX11 administration, whereas nonresponders showed an increased HP lactate-to-pyruvate ratio after the treatment. This result demonstrates the strength of the noninvasive HP-MR modality to predict treatment efficacy prior to tumor size reduction.
The β-lapachone chemotherapeutic drug acts on the quinone oxidoreductase 1 (NQO1)-mediated redox cycle, resulting in elevated superoxide and peroxide formation and in turn nicotinamide adenine dinucleotide (NAD+) depletion due to DNA damage and hyperactivation of poly(ADP-ribose) polymerase. Silvers et al [54] (

Response to Radiation Therapy
Several studies have shown that early responses to radiation therapy can be assessed using molecular imaging. Ionizing radiation generates reactive oxygen species in tumor tissues [68]. Determining oxidative stress noninvasively could measure the extent of oxidative damage. HP pyruvate-to-lactate conversion predicted the response of solid tumors to radiation therapy in animal models [56]. This is an indirect approach and exploits the fact that pyruvate-to-lactate conversion requires reducing equivalents [56]. More direct measurement of redox stress inside cells is provided by HP dehydroascorbate-based MR, as summarized in Table 1 [ [57][58][59][60][61].

Collateral Lethality
Collateral lethality is a novel therapeutic approach that exploits the deletion of passenger genes alongside neighboring (deleted) tumor suppressor genes, thus conferring cancer-specific vulnerabilities [69]. One such instance is the deletion of both copies of malic enzyme 2 (ME2) with homozygous deletion of the neighboring SMAD4 in many cases of PDAC. This makes ME3 inhibition a useful drug target because ME2 and ME3 are paralogous isoforms involved in NADPH regeneration and thus redox balance. The downstream effect of ME3 inhibition entails a reduction in the levels of branched-chain amino acid aminotransferase (BCAT) (encoded by BCAT2) via AMP-activated protein kinase-mediated mechanisms [69]. An HP α-keto isocaproate probe (Table 1), which can detect BCAT levels in vivo, could potentially be used for prognosis in the near future [66].

Imaging Peritoneal Metastasis
An interesting investigation by Eto et al [70] (Table 2) illustrates a method for the selective imaging of malignant ascites in a mouse model of peritoneal metastasis using HP-MR and bioluminescence studies [70]. In vivo HP images obtained using H 2 O and D 2 O as a radical in SUIT-2 peritoneal metastasis mice showed increasing intensity with time (0, 7, 14, and 21 days after tumor cell administration). This correlated with the increased density of bioluminescence as the density of PDAC ascites increased, thus providing the capability to monitor peritoneal metastasis as well as to evaluate the efficacy of antimetastatic drugs using these two techniques.

Metabolic Imaging Employing HP 13C Glutamine
Another possible approach for the early detection of PDAC is using HP 13

Interstitial pH Mapping
Pancreatitis (inflammation of the pancreas) and PDAC are characterized by acidic microenvironments. The interstitial pH of the pancreas is reduced in patients with chronic pancreatitis [72][73][74]. The use of pH imaging to differentiate the acidic microenvironment of pancreatic tumors from that of PanIN lesions in mice has been elucidated by Cruz-Monserrate et al [52]. Several HP probes such as bicarbonate and zymonic acid can be potentially employed to image extracellular pH in tissue, which are summarized in Table 1 [22,53,55,67,75].

Overview of AI and Deep Learning for PDAC
Deep learning has shown robust and extraordinary performance in medical image analysis. Many previous studies have explored the applications of AI, especially deep learning, for diagnosing and detecting various diseases, including pancreatic cancer, from different imaging modalities [76,77]. Leveraging HP-MR with deep learning is a promising approach to interrogate the early diagnosis and early efficacy of therapy for pancreatic cancer.
Most of the innovative applications of deep learning in biomedical imaging were triggered by convolutional neural networks (CNNs) [78], a powerful method for representation learning in images and structured data. As discussed above, neural networks, inspired by information transformation in the biological brain, require connections of all nodes of one layer to the next, which is insufficient for image analysis and fails to make use of spatial information. To overcome these issues, CNN introduces convolutional layers and pooling operations.
In addition, many innovative modifications have been proposed to boost the performance of CNN, including dropout [79], batch normalization [80], and residual learning [81]. Essentially, the input to CNN is in a grid structure to preserve the spatial information, and then multiple convolutional layers and activation layers, interspersed with pooling layers, are utilized to process the data and learn structure in each level. Furthermore, a fully connected layer computes the final outputs for image analysis tasks.
A convolutional layer includes a set of filters with learnable parameters. Each filter is slid across the width and height of the input, and the dot product of the filter and input at every special position is calculated and goes through an activation function. A nonlinear activation function, typically rectified linear units (ReLUs), expands the potential in approximation of any nonlinear function [82]. The output of a convolution layer is a stack of activation maps of all filters. For pooling layers, it takes small regions in the feature map and produces a single number as the output to extract the most significant information learned from convolutional layers.
Several variants of CNNs with innovative architectures have been proposed to achieve better performance on specific tasks or types of data. VGG [83] introduced smaller filter kernels and constructed a deeper network compared with AlexNet [84], which first utilized ReLUs, dropout, and GPU accelerations. ResNet [81] proposed residual learning by using skip connections, which not only reduces the number of parameters but also makes the network deeper at up to 152 layers without a vanishing gradient. For biomedical images, U-Net [85] constructed downstreaming and upstreaming paths for biomedical images processing, connected by a skip connection, which concatenates features to the upstreaming path. V-Net [86] extended U-Net to 3D datasets using 3D convolutional layers and achieved extraordinary performance.
To review the previous studies on using AI for PDAC, we grouped the 17 selected papers (Table 3 and Table S1 in Multimedia Appendix 1) meeting our inclusion criteria into six categories based on how AI was utilized in the context of PDAC to help recognize the gaps in the previous studies and to discuss the novel approaches that can fill the current gaps in detecting PDAC by imaging modalities.  DSC of 91% (SD 3), 89% (SD 6), 86% (SD 6), 95% (SD 2), 95% 100 patients with CT simulation scanned 3D U-Net with an attention strategy is proposed Multiorgan segmentation for pancreatic CT Liu et al [95] (SD 2), 96% (SD 1), 87% (SD 5), and 93% (SD 3) for the large bowel, small bowel, duodenum, left kidney, right kidney, liver, spinal cord, and stomach, respectively.

Tumor Growth Model
Tumor growth, especially for pancreatic neuroendocrine tumors, is related to cancer cell properties and relies on the dynamic interaction between cells and the microenvironment. Swanson et al [104] proposed a reaction-diffusion model by assuming infiltrative growth of the tumor cells but did not consider the cell metabolic rate. Liu et al [87] introduced dual-phase CT-measured intracellular volume fraction (ICVF) to the reaction-diffusion model. Cell metabolic rate was considered in the prediction of pancreatic neuroendocrine tumor growth. They evaluated the model by comparing predictions with sequential observations regarding average surface distance, root mean square deviation (RMSD) of the ICVF map, and average ICVF difference in six patients with pancreatic neuroendocrine tumors. Although the RMSD was around 4.3%, the limited number of patients involved might have undermined the final findings.

Organ/Multiorgan Segmentation and Edge Detection in Medical Images
Fu et al [88] discussed the application of a CNN consisting of 13 convolution layers and 4 pooling layers with a multilayer upsampling structure in pancreas segmentation from CT images. The proposed model was evaluated using real PDAC CT images from a dataset created by the General Surgery Department of Peking Union Medical College Hospital. The 59 patients consisted of 15 patients with nonpancreas diseases and 44 patients with pancreas-related diseases. A Dice similarity coefficient (DSC) of 76.36% was achieved. The introduced fusion layer provided good visualization for decision-making and multilayer upsampling improved the performance. However, due to the limited number of CT images for training and validation, its performance suffered from the risk of overfitting as the reported SD from precision 5-fold cross-validation was very high (mean SD of 18.08 across all classes). Moreover, the reported precision and recall for the healthy cohort (80.95 and 86.53, respectively) was much higher than that for the IPMN (75.39 and 67.37) or pancreatic neuroendocrine tumor (70.44 and 74.86) cohort.
Alternatively, to implement multiorgan segmentation, especially on abdominal CT in the pancreas, Gibson et al [89] modified V-net by replacing the convolutional layers in the encoder path by DenseNet consisting of stacks of dense blocks combined with bilinear upsampling in the decoder path. They applied this on two public datasets: one including 43 subjects from the Cancer Imaging Archive Pancreas CT data with pancreas segmentation, and the other including 47 subjects from the Beyond the Cranial Vault segmentation challenge with segmentations of all organs except the duodenum. They achieved a DSC of 78%. The introduced dense feature stack considerably reduced the number of parameters for medical image classification tasks. However, this approach is only appropriate for relatively small datasets because of overfitting issues.
As another example of an attempt to improve pancreas segmentation performance in CT scans, Boers et al [92] developed an interactive version of U-net (iUnet) by adding one interactive layer after the last fully connected layer takes feedback from annotators while freezing other layers to do the retraining. This was applied to a public CT dataset used in Gibson et al [89], which contains 90 late venous-phase abdominal CT images and a respective reference segmentation. A DSC of 78.1% (SD 8.7%) was achieved from the interactive version of iUnet, which outperformed previous methods using the same dataset. However, this approach may also suffer from overfitting issues since interactive processes may introduce external information, which limits its scalability.
Liu et al [93] presented a deep-attention U-net approach to solve multiorgan segmentation for pancreatic cancer CT images. This method achieved state-of-the-art performance, but its performance in pancreas segmentation is unclear.
Besides CT images, investigators using T1-MRI proposed several innovative approaches to segment the pancreas. Liang et al [102] introduced a top-down and bottom-up approach. In the top-down path, the initial planning contours derived from simulation MR images are transferred to daily images, and in the bottom-up path, the probabilistic support vector machine (SVM) is used with recursively retraining samples. The final result is obtained by fusing both paths and the final reported DSC was 86%. Zheng et al [103] proposed a 2D U-Net approach with shadow sets for MRI and CT pancreas segmentation. The usage of shadow sets reduced uncertainty and achieved a DSC of 84.37% on the NIH-CT-82 public dataset [105] and 73.88% on an MRI dataset collected from Changhai Hospital, including 20 patients with PDAC.

Prediction of PDAC and Risk Evaluation
To implement preoperative prediction of pancreatic neuroendocrine neoplasms (pNENs) grading by CT, Luo et al [90] applied a CNN model with identity blocks and convolution blocks to a CT imaging dataset consisting of 93 patients from the hospital. An arterial model employed for the pathological grading of pNENs achieved an area under the receiver operating characteristic curve (AUC) of 0.81. Due to the limitations of the dataset, simple deep-learning models may undermine the feature extraction ability and lead to suboptimal performance. In addition, the limited observations may lead to a lack of an independent evaluation dataset and invalidation of n-fold cross-validation, which constrains the scalability and generalizability of the proposed model. Hussein et al [101] utilized clustering and SVM with initial label estimation for risk stratification of pancreatic tumors. The model outperformed other unsupervised methods, achieving 58% accuracy out of 171 scans, of which 38 subjects were normal and the remaining 133 were diagnosed with IPMNs. In addition, Corral et al [100] performed similar experiments using MRI with CNN and SVM as the final classifier, and achieved sensitivity and specificity of 0.92 and 0.52, respectively, for the detection of dysplasia. In this task, the deep-learning protocol barely outperformed radiologists due to unbalanced data issues and the complexity of IPMN.

Diagnosis of PDAC
To implement diagnosis of PDAC from CT images, Liu et al [91] utilized a pretrained VGG16 model as a feature extraction network in conjunction with a faster recurrent CNN (R-CNN) as a decision-maker. Their CT dataset consisted of 6084 enhanced CT horizontal images from the abdomen of 338 PDAC patients. This achieved an AUC of 0.9632 for the prediction of PDAC. R-CNN usage for sequential information extraction greatly improved the diagnostic performance, and its dynamic feature extraction provided model interpretability and scalability.
Ma et al [98] utilized a regular CNN with four hidden layers on 3494 CT images from 222 patients with pathologically confirmed PDAC and 3751 CT images from 190 patients with a normal pancreas as controls from the First Affiliated Hospital, Zhejiang University School of Medicine. The overall diagnostic accuracy of trained binary classifiers was over 95%. However, it failed to beat human performance. With a similar data size, Liu et al [91] achieved better performance using a more complicated, faster R-CNN, implying the complexity of pancreatic cancer detection and its need for an appropriate model architecture design and parameter fine-tuning.
Zhang et al [99] proposed a tumor detection framework for PDAC using CT. The framework utilized feature pyramid networks with the faster R-CNN model. The Affiliated Hospital of Qingdao University provided a dataset containing 2890 CT images, and a classification AUC of 0.9455 was achieved. This framework outperformed the state-of-the-art methods, but still suffered from input uncertainty inherent in closed-source datasets. The comparison would be more effective if a public dataset was used in the experiment.

Improvement of Image Quality
Stereotactic body radiotherapy (SBRT) has shown more success in patients with locally advanced pancreatic cancer compared to conventional radiotherapy. To overcome interference due to motion in the breathing cycle and patient weight loss [106], cone-beam CT (CBCT) is commonly used for target position verification and setup displacement correction to avoid suboptimal target coverage and excessive doses to organs at risk. However, raw CBCT data cannot be used for SBRT dosage calculation due to considerable artifacts such as streaking and shading [107][108][109] caused by scatter contamination, resulting in different Hounsfield unit (HU) values from CT scans [110]. To improve the HU fidelity of CBCT, Liu et al [95] utilized self-attention cycleGan-based CBCT to synthetic CT (sCT) models on a dataset consisting of 30 patients previously treated with pancreas SBRT at Emory University. The mean absolute error of the proposed framework between CT and sCT was 56.89 (SD 13.84) HU compared to 81.06 (SD 15.86) HU between CT and the raw CBCT.

Criteria to Evaluate Annotation Accuracy in Medical Images
To test the CT data collection quality, Park et al [94] proposed two U-net-linked networks, linked by an organ-attention module, to test the performance of a well-annotated dataset, including a total of 575 participants and 1150 CT images. After appropriate management of the annotation process, an average DSC of 89.4% was achieved. This study innovatively employed a deep-learning model to test CT image annotation performance, and improved the annotation quality for further analysis and research. However, this approach still suffers from uncertainty introduced by model training and simulation.
Suman et al [97] used CNNs to train technologists in labeling pancreas segmentation CT datasets. DSC was improved through interactions between model output and expert correction, which implied that annotation quality was enhanced.

Discussion
In this review, we have discussed how two different techniques, HP-MR and AI, are revealing exciting information about PDAC and PanINs that was not accessible by diagnostic imaging even a few years ago. Deep-learning models eliminate the requirement of domain knowledge for feature engineering that is necessary for conventional machine learning models by learning from raw data. Deep-learning models are capable of learning features from the raw data and apply nonlinear transformations to map the input data to high-dimensional representations trivializing classification or regression. These models are uniquely able to transform multiple modalities into common latent space to synthesize features across all modalities to improve classification performance. However, there is no free lunch, and the flexibility and high accuracy resulting from millions of parameters comes with a requirement of a huge training dataset in comparison with other machine-learning techniques. Moreover, these models suffer from lack of interpretability and uncertainty measurement. In machine-learning algorithms, there is a tradeoff between interpretability and accuracy. When the prediction accuracy grows with more complex (increase in the number of trainable parameters) deep-learning models, the interpretability decreases. For instance, ResNet contains 5×10 7 parameters requiring 10 10 floating point operations for a single classification task, making it almost impossible to be traced or explained by humans [81,111]. Lastly, deep-learning models do not provide any uncertainty measurement to measure how certain the model is with its prediction. These models are blindly used with the assumption of "good accuracy," whereas previous experience has shown that these models are susceptible to overconfident decision-making, especially when the new data are far from the training data distribution (corner case). The lack of interpretability and uncertainty estimation is even more serious in clinical decision-making tasks since it is needed for building trust in the model's prediction.
Studies on HP-MR have demonstrated that this modality can detect metabolic changes at very early stages of lesion formation in the pancreas (eg, PanIN 1 and 2); however, this is more of a spectroscopic technique than an imaging modality. Furthermore, the signal from HP compounds lasts no more than a few minutes depending on the T 1 that allows for rapid acquisition of dynamic metabolic flux in the organ of interest. Table 4 summarizes the strengths and weaknesses of AI, MRI, and HP-MR. Table 4. Strengths and weaknesses of artificial intelligence (AI), magnetic resonance imaging (MRI), and hyperpolarized magnetic resonance (HP-MR).

Weaknesses Strengths Technique
Poor signal-to-noise ratio and contrast-to-noise ratio.
Cannot detect pancreatic cancer at early stages.

Rapid acquisition of anatomical images.
Well-established and widely distributed imaging modality.

MRI
Short time window of imaging (~2 minutes).
Expensive initial investment in the infrastructure.
Slow adoption in the clinical setting.
Real-time metabolic flux measurements at the organ of interest.
Can detect premalignant stages of pancreatic cancer.

HP-MR
Intensive data requirement.
High uncertainty on corner cases.

Lack of interpretability.
No feature engineering, ability to learn features from raw data.
Ability to learn features from and across multiple modalities.
High accuracy result.

AI
To take advantage of the strengths of AI, MRI, and HP-MR, and mitigate their weaknesses, we propose the following pipeline as illustrated in Figure 4. Our pipeline leverages the unique capability of AI to learn features from each and across both HP-MR and MRI as complementary modalities to investigate the early detection of PDAC by overlaying the anatomical imaging for localized spectroscopic information of real-time metabolic flux in the pancreas. Additionally, we utilize Grad-CAM [112,113] and concrete dropout to provide a visual explanation, and introduce Bayesian inference to estimate uncertainty in the model's decision. The training process of our pipeline is as follows: axial, sagittal, and coronal MR images in the T1 and T2 modalities are annotated to highlight the pancreas area by radiologists to train a deep-learning semantic segmentation network developed by our team. We extract the ROIs from MR images (ie, the pancreas). The extracted ROIs with metabolic information from HP-MR are the inputs for our multimodal deep-learning model to predict pancreatic cancer status. The appropriate combination of MRI and HP-MR as complementary modalities improves the classification performance. Therefore, the ground truth for our second deep-learning model is the presence of early stages of PDAC established by pathology reports and electronic health records of the patients. The training path is shown with the dashed lines and the inference path is shown with the solid lines in Figure 4. It has been estimated that there is a window of opportunity of ~10 years from the moment in which a pancreatic epithelial cell undergoes an oncogenic hit and the time of diagnosis of, often fatal, pancreatic cancer [46,114]. Together, AI, HP-MR, and conventional MRI as complementary modalities can address this knowledge gap in diagnostic imaging within this crucial time window of opportunity to save lives.
Leveraging AI and HP-MR applications together may lead to the development of real-time actionable biomarkers of early detection, assessing aggressiveness, and interrogating the early efficacy of therapy in PDAC. For example, multimodal AI can learn features from both HP-MR, as well as anatomical MRI and CT imaging modalities, to yield "hybrid biomarkers" and reduce the time required to detect PDAC evolution in three key areas of tumor progression: initial development of the tumor, its regression following therapy, and the eventual recurrence of the tumor. This innovative synthesis of these techniques may result in a more sensitive readout of tumor progression that can be readily translated and significantly impact how PDAC patients, as well as patients at high risk of developing this deadly disease, are currently managed in the clinic.